E2B says NVIDIA Vera can power up to 1,550 loaded sandboxes

E2B says NVIDIA Vera delivered faster single-thread compute and memory population in its tests, while supporting up to 1,550 loaded sandboxes on one server. The company is evaluating the platform as a foundation for agent workloads that need many isolated computing environments.
- Vera delivered 1.51× single-thread compute throughput versus E2B's AMD reference.
- Memory population reached 62.8 GB/s, compared with 17.4 GB/s on AMD.
- One tuned Vera server sustained 1,400 to 1,550 loaded sandboxes.
- E2B measured 50 ms warm starts and 3.8 ms low-load command round trips.
- The company completed 22,145 create-and-destroy cycles with zero errors.
Vera targets the needs of agent sessions
E2B provides isolated computers through an SDK, using microVM-based sandboxes for tasks such as repository searches, file edits, test runs and scripts. The company says its runtime has no cold start because each sandbox resumes from a snapshot, with memory loaded only as needed.
That design makes processor and memory performance important even when an AI model runs elsewhere. An agent may need the output of one command before it can choose the next, while many other sandboxes can be installing packages, running scripts or waiting for a model response.
According to E2B, NVIDIA Vera fits this workload through fast compute, memory bandwidth for snapshot restores and capacity for many sandboxes on a single server.
Benchmark results show gains in compute and memory population
In E2B's single-thread compute test, Vera delivered 7,759 events per second, compared with 5,131 on the company's AMD EPYC 9B45 reference. E2B reports this as 1.51× the throughput, or 1.5× in its headline result.
For memory population using the same 2 MiB huge-page backing as its sandboxes, Vera reached 62.8 GB/s. The AMD reference reached 17.4 GB/s, giving Vera 3.61× the throughput, according to E2B.
E2B cautions that the memory test measures memory population rather than complete sandbox startup, and says the performance of individual tools and applications will vary. The company connects the result to its snapshot-based runtime, where memory is populated as environments resume and run.
One server sustained up to 1,550 loaded sandboxes
E2B tested its single-node E2B Embed distribution directly on Vera bare metal. In separate tests, the company measured 50 ms warm, sequential sandbox starts and 3.8 ms command round trips at low load. E2B says these figures are not fleet-load latencies.
For a mixed workload, each sandbox ran a repeating cycle of a CPU burst, a 10 MiB disk write and a pause, while one in ten ran continuously. Across four tuned runs, the server sustained 1,400 to 1,550 loaded sandboxes within the test's latency and error objectives. At 1,550 sandboxes, command round-trip latency was 255 ms at the 95th percentile. A workload in which every sandbox ran continuously reached 550.
Configuration changes were also significant. E2B says initial runs reached 320 sandboxes because of network and device limits, while host tuning raised the result to as many as 1,550. In a separate churn campaign, the company completed 22,145 create-and-destroy cycles with zero errors.
E2B says its next steps include a multi-node Vera cluster, cold resumes backed by object storage and an x86 comparison using the same E2B packaging and host tuning. For developers, the reported results point to faster individual environments and the ability to run many agent attempts in parallel, although the tests were conducted on one pre-production dual-socket server.


